A local color descriptor for efficient scene-object recognition

This paper presents an effective use of local descriptors for object or scene recognition and indexing. This approach is in keeping with model-based recognition systems and consists of an extension of a standard point-to-point matching between two images. Aiming at this, we address the use of Full-Zernike moments as a reliable local characterization of the image signal. A fundamental characteristic of the used descriptors is then their ability to "absorb" a given set of potential image modifications. Their design calls principally for the theory of invariants. A built-in invariance to similarities allows one to manage narrow bounded perspective transformations. Moreover we provide a study of the substantial and costless contribution of the use of color information. In order to achieve photometric invariance, different types of normalization are evaluated through a model-based object recognition task.